Texture Feature Recognition Based On NSCT And Support Vector Machine

The texture characteristic is quantized representation of the image internal graylevel changes. The image texture analysis and recognition is one of the importantresearch content of the image processing and has broad application prospects. Inrecent years, many scholars have done a lot of research on texture analysis and havemade a number of significant achievements. However, the texture analysis is still acomplex problem and required in-depth study. Texture feature extraction is the basicof the texture analysis. It determines the similarity between the samples and is the keyof designing classifier. How to extract the less characteristic dimension and thecalculation is simple, at the same time be able to represent the image texture featuresoptimal, is the emphasis and difficulty in the texture feature extraction. Based onin-depth study of texture feature extraction method, in this thesis, nonsubsampledcontourlet transform and support vector machine are combined to use for imagetexture feature extraction and recognition. The optimal feature vectors whichrepresent the image texture accurately when the classifier is SVM was found. Thismethod is applied to the study of medical CT image recognition.1In-depth study the image texture feature extraction methods. The image wastransformed from the space domain to the frequency domain by the nonsubsampledcontourlet transform. The principles of the contourlet transform and nonsubsampledcontourlet transform was explains detailed.2Each sub-band image texture features were extracted based on nonsubsampledcontourlet transform. The combination of feature vectors which can reflect the imagetexture features were found and select the feature vector and dimensionality reductionprocessing. The appropriate parameters were select to establish a support vectormachine model. The extracted feature vectors were input to support vector machinesfor classification. Brodatz database was used for simulation. Results demonstrate thatmean and variance of low frequency and the energy of high frequency are combined to make the recognition accuracy higher. Moreover, the vector dimension is less andthe recognition is faster. These features are the optimal representation of the imagetexture when the classifier is SVM.3In connection with the characteristic of brain CT image, nonsubsampledcontourlet transform and support vector machine were combined for medical CTimage classification and recognition. Brain CT images were decomposed throughNSCT. The low-frequency sub-band mean and variance and high-frequency sub-bandenergy were extracted to form vectors group as brain CT image texture features andinput to SVM for classification. The experiments show that the features extracted inthis article make the non-diseased brain CT image recognition accuracy rates up to96.67%and the diseased brain CT image recognition rates up to90%. The disease andnon-diseased brain CT image can be distinguished relatively effective.